无重叠视域多摄像机目标跟踪系统若干关键问题研究

In recent years, video surveillance systems have been developing towards the large-scale and intelligent direction. Non-overlapping multi-camera tracking is one of the fundamental problems in constructing large-scale and intelligent systems. We focus our research on some key issues of the non-overlapping multi-camera tracking in the proposal. In general, we present a non-overlapping multi-camera tracking framework based on the Bayes theory, as well as a computing model by combing constraints of object appearance and camera topology. Our research proposal is mainly composed of the following aspects. First, we study the problem of multi-feature object representation and propose a color-transfer model based on an object reflection map, where illumination changes are automatically adapted. Second, we study the problem of estimating camera topologies and propose a weekly supervised online learning method to automatically obtain the topologies as well as their variations. Third, the object association issue based on multi- feature is studied. To emphasize the contributions of different features, an object association model is proposed, which combines descriptive model and multi-kernel based discriminative model. Finally, aiming at obtaining a global optimum solution to the object association problem, we also present a constrained discrete particle swarm optimization method, in order to speed up solving the multi-object optimal path problem. The research on the aforementioned issues is expected to make contributions to the aspects such as intelligence enhancement for video surveillance systems, discipline construction enrichment, and thereof technical support for the social security.

近年来，视频监控系统正在向大型化、智能化方向发展。无重叠视域多摄像机目标跟踪是大型智能视频监控系统的核心问题之一。项目针对无重叠视域多摄像机目标跟踪中的若干关键问题开展研究，提出一种基于贝叶斯理论的无重叠视域多摄像机目标跟踪框架，并且给出目标表现模型约束和摄像机拓扑关系约束相结合的计算模型。主要研究内容包括：①研究多特征目标表达问题，提出一种基于目标反射图的颜色转移模型，该模型能自动适应光照变化。②研究摄像机拓扑关系估计问题，提出一种能适应拓扑关系动态变化的弱监督在线学习方法，自动获取拓扑关系。③研究多特征目标关联问题，提出一种描述型关联和多核学习判别型关联相结合的关联模型，以体现不同特征的作用。④研究全局优化目标关联问题，提出一种带约束优化的离散粒子群算法，快速求解多目标最优路径问题。研究成果在提高视频监控系统的智能性、丰富学科建设和保障社会安全等方面具有重要意义。

近年来，视频监控系统正在向大型化、智能化方向发展。无重叠视域多摄像机目标跟踪是大型智能视频监控系统的核心问题之一。项目针对无重叠视域多摄像机目标跟踪中的若干关键问题开展研究，提出一种基于贝叶斯理论的无重叠视域多摄像机目标跟踪框架，并且给出目标表现模型约束和摄像机拓扑关系约束相结合的计算模型。主要研究内容包括：①研究了多特征表达问题。研究主要包括：传统特征（低层特征以及这些特征的融合、显著性特征、语义属性特征和协方差描述符）、深度特征。在本项目的研究过程中分别对这些特征进行了改进和创新，并提出多种行人图像分割的方法。②研究了摄像机拓扑关系估计问题，提出利用混合型高斯模型估计摄像机进出口分布，利用目标之间的相似度计算得到转移时间概率，然后根据学习得到的转移时间概率分布判断进出口的路径关系。③研究了相邻摄像机间的目标关联，提出了基于测度学习的目标匹配和基于贝叶斯框架的目标关联，即融合目标外观特征和摄像机拓扑关系实现目标匹配。④研究全局优化目标关联问题，提出多种但摄像机多目标跟踪算法。主要方法有：基于广义关联聚类图的分层关联多目标跟踪、基于运动动态和多层超图关联的多目标跟踪、改进的离散连续优化多目标跟踪、基于自适应压缩特征选择的实时目标跟踪算法，有效提升了跟踪准确率。

内容获取失败，请点击重试